Bayesian neural network with autoencoder for model-based description of $α$-particle preformation factor

This study introduces a hybrid Bayesian neural network with an autoencoder framework integrated with the cosh potential to significantly improve the prediction accuracy and uncertainty quantification of the $\alpha$-particle preformation factor, successfully revealing nuclear shell effects and validating the stability of the predicted superheavy island near $N=184$.

The Gist

Imagine the atomic nucleus as a crowded, chaotic dance floor. Inside this dance floor, particles called protons and neutrons are constantly moving. Sometimes, four of them (two protons and two neutrons) decide to huddle together, form a tight little group called an alpha particle, and try to escape the dance floor entirely. This escape is what we call alpha decay.

The big question physicists have always struggled with is: How likely is it that these four dancers will actually form a group and break free?

In the world of physics, this likelihood is called the preformation factor. Think of it as the "grouping probability." If the probability is high, the nucleus is unstable and decays quickly. If it's low, the nucleus hangs around longer.

The Old Way: Guessing with a Ruler

For decades, scientists tried to calculate this probability using complex mathematical formulas (like the "CPT-LSM" mentioned in the paper). It was like trying to predict the weather by looking at a single thermometer and a ruler.

• The Problem: These old formulas worked okay for some nuclei, but they were rigid. They couldn't account for the weird, messy details of how particles interact. They often missed the "odd-even" patterns (like how pairs of dancers are more stable than single dancers) and the "shell effects" (like how a full dance floor behaves differently than a half-empty one).
• The Result: Their predictions were often off by a wide margin, especially for the heaviest, most unstable elements.

The New Way: The "Smart Detective" (BNN-Auto)

This paper introduces a new, super-smart detective team to solve the mystery. They built a hybrid machine learning model called BNN-Auto. Let's break down what that means using a simple analogy:

1. The Autoencoder (The "Compression Artist"):
   Imagine you have a massive library of nuclear data. It's messy and full of details. The "Autoencoder" is like a brilliant librarian who can read a thousand books, summarize the most important plot points into a single, perfect sentence, and then reconstruct the story perfectly from that sentence. It learns to strip away the noise and find the true essence of what makes a nucleus stable or unstable.

2. The Bayesian Neural Network (The "Uncertainty-Aware Detective"):
   Most computer programs are like overconfident students who give you an answer and say, "I'm 100% right!" even if they are guessing.
   The Bayesian part of this model is different. It's like a detective who says, "Based on the clues, I think the answer is X, but I'm only 80% sure. Here is the range of possibilities."

   • Why this matters: In science, knowing how sure you are is just as important as the answer itself. This model doesn't just predict the decay time; it tells you how much it trusts its own prediction.

What Did They Discover?

By feeding this "Smart Detective" data from 535 different nuclei, the team achieved some amazing results:

• Massive Accuracy Boost: The new model reduced the error in its predictions by over 50% compared to the old methods. It's like going from a weather forecast that gets the temperature wrong by 20 degrees to one that's off by only 5 degrees.
• Spotting the Patterns: The model naturally "learned" the hidden rules of the universe without being explicitly told them.
  • The "Odd-Even" Staggering: It correctly predicted that nuclei with even numbers of protons and neutrons are much more stable (like a happy couple) than those with odd numbers (like a lone wolf).
  • The "Shell" Effects: It noticed that when neutron numbers hit specific "magic numbers" (like 126 or 184), the nucleus becomes extra stable, almost like a fortress.
• Predicting the Future (The Island of Stability): The team used their model to predict the behavior of super-heavy elements (specifically element 120) that haven't been fully studied yet.
  • The Big Reveal: They predicted that near a neutron number of 184, there is an "Island of Stability." Just like a calm island in a stormy ocean, these super-heavy nuclei might actually last longer than their neighbors. This gives experimentalists a target to aim for when building new elements.

The Bottom Line

This paper is a game-changer because it combines the best of two worlds: the deep physical laws of nuclear structure and the pattern-recognition power of modern AI.

Instead of just forcing nature into a rigid mathematical box, they built a flexible, learning system that can "feel" the nuances of the atomic dance floor. It doesn't just give us better numbers; it gives us a clearer map to explore the heaviest, most mysterious elements in the universe, potentially guiding us to the "Holy Grail" of super-heavy, long-lasting elements.

Technical Summary

Here is a detailed technical summary of the paper "Bayesian neural network with autoencoder for model-based description of α-particle preformation factor."

1. Problem Statement

The study addresses the challenge of accurately describing and predicting the $\alpha$-particle preformation factor ($P_\alpha$), a critical physical quantity in determining $\alpha$-decay half-lives for heavy and superheavy nuclei.

• Theoretical Limitations: Traditional microscopic calculations of $P_\alpha$ are extremely complex and rely heavily on model assumptions. Phenomenological approaches often treat $P_\alpha$ as a constant or use semi-empirical parameterizations (e.g., the CPT-LSM model), which suffer from limited generalizability, especially in the superheavy region far from stability.
• Data Scarcity & Uncertainty: Existing machine learning applications in nuclear physics often lack uncertainty quantification and tend to overfit when dealing with limited datasets (e.g., 535 nuclei in this study).
• Goal: To develop a robust, data-driven framework that can adaptively learn systematic nuclear structure trends, provide uncertainty estimates, and accurately predict $P_\alpha$ and $\alpha$-decay half-lives for known and unknown nuclei.

2. Methodology

The authors propose a hybrid framework integrating Bayesian Neural Networks (BNN) with an Autoencoder (Auto) architecture, grounded in the cosh potential (CPT) model for $\alpha$-decay.

A. Theoretical Foundation

• $\alpha$-Decay Model: The decay width ($\Gamma$) and half-life ($T_{1/2}$) are calculated using the WKB approximation with a cosh potential for the nuclear interaction. The preformation factor is extracted inversely from experimental data using the relation:
  $$P_\alpha^{exp} = \frac{T_{1/2}^{cal}}{T_{1/2}^{exp}}$$
  where $T_{1/2}^{cal}$ assumes $P_\alpha = 1$.

B. The BNN-Auto Architecture

• Input Features: The model takes 8 features per nucleus: Mass number ($A$), Neutron number ($N$), Proton number ($Z$), Decay energy ($Q_\alpha$), Orbital angular momentum ($l$), and three derived features ($\lambda_1 = A^{1/3}$, $\lambda_2 = Z/\sqrt{Q_\alpha}$, $\lambda_3 = \sqrt{l(l+1)}$).
• Network Structure:
  • Encoder: Three fully connected layers with Batch Normalization and ReLU activation to compress input features into a latent space.
  • Decoder: Symmetrically reconstructs the input using a Sigmoid activation to ensure robust feature representation.
  • Prediction Head: Maps latent variables to the final $P_\alpha$ prediction using an ELU activation.
• Bayesian Inference:
  • Network weights are treated as random variables with a zero-mean Gaussian prior.
  • Variational Inference (VI) is employed to approximate the posterior distribution of weights, minimizing the Kullback-Leibler (KL) divergence.
  • Loss Function: A composite loss function balances three terms:
    $$L = k L_{recon} + (1-k) L_{pre} + KL(q(\omega|\theta) \| p(\omega))$$
    Where $L_{recon}$ is the autoencoder reconstruction error, $L_{pre}$ is the prediction error (negative log-likelihood), and $KL$ is the regularization term.
• Training: Optimized using AdamW with 10-fold cross-validation repeated 5 times over 1000 epochs.

3. Key Contributions

1. Hybrid Framework: First integration of BNN with Autoencoder specifically for $\alpha$-particle preformation, combining the uncertainty quantification of Bayesian methods with the feature robustness of autoencoders.
2. Uncertainty Quantification: Unlike deterministic neural networks, the BNN provides posterior distributions for predictions, allowing for the calculation of mean values and standard deviations (uncertainty estimates) via Monte Carlo sampling.
3. Systematic Bias Correction: The model effectively identifies and corrects systematic biases in traditional theoretical models (like CPT-LSM) related to nuclear shell closures and odd-even effects.
4. Superheavy Prediction: Successfully applied the model to predict $\alpha$-decay properties for unknown superheavy nuclei with $Z=120$, verifying the "island of stability" near $N=184$.

4. Results

The model was trained and validated on a dataset of 535 nuclei (159 even-even, 295 odd-A, 81 odd-odd).

• Performance Metrics (Root Mean Square Deviation, $\sigma_{RMS}$):

  • The BNN-Auto method significantly outperformed the traditional CPT-LSM model.
  • Training Set: Relative improvement in $\sigma_{RMS}$ of 61.14% for even-even nuclei and 30.75% overall.
  • Validation Set: Relative improvement of 54.49% for even-even nuclei and 35.37% overall.
  • Odd-Odd Nuclei: Showed the highest relative improvement on the validation set (66.74%), demonstrating strong generalization to complex nuclear structures.

• Physical Insights:

  • Odd-Even Staggering: The model clearly reproduced the higher $P_\alpha$ and shorter half-lives for even-even nuclei compared to odd-A and odd-odd nuclei, reflecting pairing correlations.
  • Shell Effects: Pronounced shell effects were observed near magic numbers ($N=126$, $N=150-154$, $Z=82$). Specifically, a suppression of $P_\alpha$ was noted near closed shells, leading to longer half-lives.
  • Residual Analysis: The residual distribution of the BNN-Auto method was significantly narrower (Std dev reduced from 0.418 to 0.280) and centered closer to zero compared to the CPT-LSM method, indicating a uniform enhancement in accuracy across the nuclide chart.

• Superheavy Predictions ($Z=120$):

  • Using $Q_\alpha$ values from the WS4+ mass model, the BNN-Auto predicted half-lives for $Z=120$ isotopes.
  • A significant increase in half-life was observed near $N=184$, confirming the shell stabilization effect associated with the predicted "island of stability."

5. Significance

• Theoretical Advancement: This work provides a high-precision theoretical description of $\alpha$-decay that overcomes the limitations of parameterized models by learning directly from data while respecting physical constraints.
• Reliability: The inclusion of uncertainty quantification makes the predictions more reliable for guiding future experiments, particularly in regions where experimental data is scarce (superheavy elements).
• New Paradigm: It establishes a new data-driven paradigm for exploring nuclear structure evolution under extreme conditions, proving that machine learning can effectively extract microscopic nuclear structure information (like shell closures and pairing) from macroscopic decay observables.
• Experimental Guidance: The specific predictions for $Z=120$ nuclei provide concrete targets and theoretical benchmarks for upcoming experiments aiming to synthesize and characterize superheavy elements.